forecaster stringclasses 10
values | forecast_date stringclasses 5
values | horizon_year int64 2.03k 2.04k | scope stringclasses 2
values | metric stringclasses 3
values | unit stringclasses 3
values | value_low int64 78 1.07k | value_high int64 78 1.07k | revised_value float64 106 220 ⌀ | revised_date stringclasses 2
values | transparency stringclasses 3
values | verified stringclasses 2
values | source stringlengths 29 111 | notes stringlengths 107 231 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
IEA | 2025 | 2,030 | Global | demand | TWh | 946 | 946 | null | null | open | yes | IEA Energy&AI annex World Data r24 | Base Case (2024=416); IEA OWN 2030 scenarios span 669 Headwinds-1264 Lift-Off; quoted 945 is base-case-only |
Gartner | 2025-11 | 2,030 | Global | demand | TWh | 980 | 980 | null | null | opaque | yes | Gartner newsroom press release 2025-11-17 | CONFIRMED 448 TWh(2025)->980(2030) Global; AI servers 93->432 TWh; transparency=opaque means model proprietary/not reproducible (release is free but unauditable) |
McKinsey | 2024 | 2,030 | US | demand | TWh | 606 | 606 | null | null | opaque | yes | McKinsey "How data centers...sate AI's hunger for power" (17 Sep 2024) + "AI's power binge" charts (6 Nov 2024) | CONFIRMED at McKinsey primary: 147(2023/3.7%)->224(2025/5.2%)->606(2030/11.7%); medium scenario full yr-by-yr; ~23% CAGR; figure PRIMARY but model proprietary/not reproducible |
LBNL | 2024 | 2,028 | US | demand | TWh | 325 | 580 | null | null | open | yes | LBNL 2024 US Data Center Energy Usage Report (Shehabi et al; eta-publications.lbl.gov) | 325-580 TWh = 6.7-12% of US elec in 2028 NOT 2030 (common misquote; CSV year corrected 2026-06-06); 2023 actual=176 TWh/4.4%; the US-govt anchor |
EPRI | 2024 | 2,030 | US | demand | TWh | 200 | 400 | null | null | open | yes | EPRI Powering Intelligence (May 2024; restservice.epri.com) | PRIMARY UNIT is % not TWh: scenarios 4.6/5.0/6.8/9.1% of US elec by 2030 (TWh DERIVED/approx); EPRI 2026 ed REVISED UP to 9-17% by 2030 (~+60%) = self-revision datapoint |
BCG | 2024 | 2,030 | US | demand | TWh | 1,050 | 1,050 | null | null | opaque | no | via WRI (NOT the in-hand BCG PDF) | CHECKED in-folder BCG "Infrastructure Strategy 2026" p20: gives only a GROWTH RATE (compute demand high-teens pct/yr to 2030 vs hist 11-12pct) NOT 1050 TWh; the 1050 is from a SEPARATE paywalled BCG study -> stays opaque/unverified |
GoldmanSachs | 2025-02 | 2,030 | Global | growth | pct_vs_2023 | 165 | 165 | 220 | 2026-04 | opaque | yes | Goldman Sachs (Feb 2025 "165% by 2030" -> GS SUSTAIN Nov 2025 "175%" -> Apr 2026 "~220%") | CONFIRMED DOUBLE up-revision 165->175->220 pct vs2023 (all primary GS); ~1350 TWh global equiv per secondary; model proprietary |
SP_Global | 2025 | 2,030 | US | demand | TWh | 728 | 728 | null | null | secondary | yes | S&P Global 451 Research Market Monitor (Sep 2025; spglobal.com news) | base 366 TWh(2025)->728(2030) US; HYPERSCALE/LEASED/CRYPTO ONLY (excl enterprise=scope caveat); also 61.8->134.4 GW; "nearly triple" is GW-from-2024 (~2x in TWh) - unit nuance |
BloombergNEF | 2025-04 | 2,035 | US | capacity | GW | 78 | 78 | 106 | 2025-12 | secondary | yes | BNEF "AI and the Power Grid" (1 Dec 2025; via Utility Dive/Bloomberg) | REVISED UP 78(Apr25)->106(Dec25) GW =+36% in 7mo; 2035 horizon + GW (not TWh); BNEF calls 106 CONSERVATIVE vs GS/BCG/McKinsey; TX 12GW early-stage only 1.8GW location-confirmed (speculative) |
McKinsey | 2024 | 2,030 | Global | capacity | GW | 220 | 220 | null | null | opaque | yes | McKinsey "Scaling bigger faster cheaper data centers" (Exhibit 2) | GLOBAL ~220 GW by 2030 (capacity); SAME forecaster reports US in TWh (606) but GLOBAL in GW = intra-forecaster unit-switch; NOT directly comparable to IEA/Gartner global-TWh (no clean conversion) |
Deloitte | 2025 | 2,030 | Global | demand | TWh | 1,065 | 1,065 | null | null | secondary | yes | Deloitte 2025 TMT Predictions | 536 TWh(2025/~2pct)->1065(2030/~4pct); scenario range ~1000-1300; built on EIA IEO-2023 base + named sources (validated) = methodology sketched (more transparent than peer consultancies) |
AI Energy-Demand Forecast Scorecard
A reproducible audit of how the field forecasts data-centre electricity demand: how the published forecasts disperse, how they get revised, and whether they are transparent enough to reproduce. Primary-sourced, published with the data and a script that regenerates every figure.
- Author: NM AI Research (independent analyst)
- ORCID: 0009-0003-4213-7769
- DOI: https://doi.org/10.5281/zenodo.20572928
- Interactive tool: https://nmairesearch.github.io/forecast-scorecard/
- Source and code: https://github.com/NMAIResearch/forecast-scorecard
Files
forecast_scorecard_data.csv(11 rows): one row per published forecast. Columns:forecaster,forecast_date,horizon_year,scope,metric,unit,value_low,value_high,revised_value,revised_date,transparency,verified,source,notes.build.py: standard-library reproducer that reads the data and writes the front-end.LICENSE: Creative Commons Attribution 4.0 International.
Method
Dispersion is measured only within comparable slices, because units and scopes are not interchangeable. Each forecast is traced through a transparency funnel from verified to confirmable to reproducible. Drafting is AI-assisted; the judgement is not.
Citation
NM AI Research. AI Energy-Demand Forecast Scorecard. Zenodo. https://doi.org/10.5281/zenodo.20572928 . Licensed CC BY 4.0.
- Downloads last month
- 44